Capacity-Stability Trade-off

• Capacity–stability trade-off is a phenomenon where systems gain higher memory, throughput, or density at the expense of robustness, retention, or recoverability.
• The trade-off appears in diverse fields such as memristors, echo state networks, distributed storage, and graph neural networks, each balancing usable degrees of freedom against mechanisms that limit drift or failure.
• Analytical models and experimental analyses demonstrate that optimizing capacity involves costs in energy, error correction, and contractive dynamics, underscoring a fundamental balance in design.

In recent literature, the capacity–stability trade-off denotes a family of constraints in which greater storage density, memory depth, representational flexibility, throughput, or information-integration ability is obtained at the price of weaker retention, robustness, contraction, or recoverability. The precise meanings of capacity and stability are domain-specific, but the recurring structure is that useful information-carrying degrees of freedom compete with mechanisms that suppress drift, perturbation growth, or failure accumulation (El-Geresy et al., 22 Aug 2025, Singh et al., 24 Jul 2025, Gao et al., 2023, Luby, 2016, Yi et al., 30 Dec 2025).

1. Formal meanings of capacity and stability

Across the cited works, capacity is operationalized in several non-equivalent but structurally related ways. In Self-Directed Channel memristors, capacity is the mutual information of a delay-dependent storage channel under an average programming-energy constraint,

$C_E(d)=\max_{P(R)\in\Gamma_B} I(R;\hat R),$

where the input is a programmed resistance state $R$, the output is the delayed readout $\hat R$, and $\Gamma_B$ enforces the energy budget (El-Geresy et al., 22 Aug 2025). In Echo-State Networks, capacity is the delay-resolved memory capacity

$$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$

where $n$ is reservoir dimension; stability is formalized by the Echo-State Property and the Fading-Memory Property (Singh et al., 24 Jul 2025). In graph neural networks, representational capacity is proxied by degrees of freedom in the parameter space, whereas stability is a Lipschitz-type robustness of the output map $\Phi_\Theta(S,x)$ to topological perturbations of the graph shift operator $S$ (Gao et al., 2023). In distributed storage, capacity is the amount of source data that remains recoverable for long durations under continual failures and repairs, asymptotically bounded by

$$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$

with $E$ the erasure rate and $R$0 the repair read rate (Luby, 2016, Luby et al., 2021). In optical quantum memory, capacity is linked to the maximal photon-number variance supported by the memory and is approximated by

$R$1

while stability is high-fidelity state preservation over storage and driving times (Yi et al., 30 Dec 2025).

The stability side is equally heterogeneous. In dynamical systems it appears as contraction, negative Lyapunov exponents, or monotone decay of a Lyapunov functional (Singh et al., 24 Jul 2025, Ito, 2019). In storage devices it appears as retention against drift toward equilibrium (El-Geresy et al., 22 Aug 2025). In communication and infrastructure it appears as recoverability under node failures, robustness of a min-cut under variable link capacities, or protection of a primary user under interference constraints (Luby, 2016, Gitik et al., 20 Jul 2025, Yazdani et al., 2018). This suggests a unifying pattern: capacity is always computed under an explicit channel, dynamical, or recoverability constraint, and stability determines how sharply that feasible set contracts.

2. Physical memories and retention-limited storage

In SDC memristors, the trade-off is formulated as a capacity–stability–energy triangle. The memristive state is a resistance-like variable $R$2, and retention is modeled as a delay-dependent noisy channel

$R$3

with all states drifting toward an equilibrium resistance $R$4. The programming cost from equilibrium is fit by

$R$5

with fitted parameters $R$6 and $R$7. Using cGAN-generated channel estimates and cost-constrained Blahut–Arimoto optimization over $R$8 input and output bins, the work reports concave energy–capacity curves for 10, 50, and 100 minute delays, with $R$9 at any given energy, and with capacity saturating smoothly rather than abruptly at high energy (El-Geresy et al., 22 Aug 2025). The factual implication is that stronger retention requirements collapse distinguishable levels over time even when more programming energy is available.

The optical quantum-memory setting yields a sharper analytical statement. In an EIT-based ensemble memory with disordered detuning $\hat R$0 and disordered coupling $\hat R$1, the stored state acquires a number-dependent Berry phase. The paper shows that disordered coupling contributes to decoherence only through its correlation with detuning, not as an isolated imperfection. Averaging over disorder gives the fidelity lower bound

$\hat R$2

and, after relating $\hat R$3 to the LSD capacity $\hat R$4, the high-reliability regime obeys

$\hat R$5

Here $\hat R$6 is storage time, $\hat R$7 is driving time, and $\hat R$8 depends only on the pulse shape. The result makes the trade-off explicit: at fixed disorder and atom number, raising capacity forces shorter storage or driving times if fidelity is to remain high (Yi et al., 30 Dec 2025).

A common misconception is that all physical imperfections contribute independently. The quantum-memory analysis rejects that view: coupling disorder alone does not decohere the dark-state protocol when $\hat R$9; the harmful term is the correlated $\Gamma_B$0 contribution to the random Berry phase (Yi et al., 30 Dec 2025). In memristors, similarly, the limiting factor is not only programming noise but the convergence of delay-dependent output distributions toward a common equilibrium distribution (El-Geresy et al., 22 Aug 2025).

3. Dynamical computation and machine learning

In Echo-State Networks, the trade-off is stated in dynamical-systems language. Stability is formalized by the Echo-State Property, which requires wash-out of initial conditions, and by the Fading-Memory Property, which requires geometric forgetting of remote inputs. A sufficient contraction condition is

$\Gamma_B$1

where $\Gamma_B$2 is the activation Lipschitz constant and $\Gamma_B$3 is the reservoir matrix. Capacity is quantified by the memory-capacity spectrum $\Gamma_B$4 and the total bound $\Gamma_B$5. The paper identifies three regimes: too contractive ($\Gamma_B$6), where memory is shallow and concentrated at short delays; too expansive, where ESP can fail and positive Lyapunov exponents make memory irreproducible; and a near-critical regime with $\Gamma_B$7, where ESP and FMP still hold but forgetting is slow enough to spread capacity over longer delays (Singh et al., 24 Jul 2025). The trade-off is therefore not between stability and instability in the abstract, but between useful long-range memory and the contraction margin required for reproducible computation.

Continual learning gives a parameter-space version of the same dilemma. Auxiliary Network Continual Learning introduces a frozen old network $\Gamma_B$8, a plastic auxiliary network $\Gamma_B$9 trained only on the current task, and a main network trained with

$$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$0

For EWC- and MAS-type regularizers, the resulting optimum is an interpolated parameter target between old and auxiliary solutions, weighted by $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$1 and $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$2. For LFL and LwF, the interpolation appears in feature or logit space. The paper’s analyses with weight distance and CKA show that tuning $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$3 moves the main solution continuously between stability and plasticity, and that the best-performing solutions are those whose representations are closest to the multitask solution rather than to either extreme. On the reported benchmarks, ANCL improves strong baselines by roughly $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$4–$$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$5 absolute in both task-incremental and class-incremental settings (Kim et al., 2023).

These two cases expose complementary mechanisms. In ESNs, stronger stability consumes dynamical degrees of freedom by enforcing contraction (Singh et al., 24 Jul 2025). In continual learning, stronger stability consumes parameter degrees of freedom by preventing reuse of weights needed for new tasks (Kim et al., 2023). In both, the highest utility lies neither at maximal rigidity nor at unconstrained plasticity.

4. Graph structure, higher-order organization, and criticality

For graph neural networks, the trade-off is formulated between representational capacity and robustness to graph perturbations. The EdgeNet framework unifies more than twenty architectures, including convolutional, attention-based, graph isomorphism, and hybrid models. Capacity is proxied by parameter-space degrees of freedom: classical graph convolutions use one scalar per shift, node-varying models use one parameter per node per hop, and fully edge-varying models use one parameter per edge per hop. Stability is expressed through bounds of the form

$$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$6

for relative topological perturbations $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$7. The key mechanism is eigenvector misalignment between parameter matrices and the graph shift operator. The stability constants obey

$$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$8

so more flexible and less spectrally aligned architectures are less stable. This makes the trade-off quantitative: greater representational freedom enlarges the space of realizable graph operators but worsens transferability under topology change (Gao et al., 2023).

A different but related formulation appears in evolved Boolean networks. There, the paper contrasts redundancy-dominated, synergy-dominated, and Tononi–Sporns–Edelman-complex networks. Capacity is the integrated-information quantity $$C(\tau)=1-\frac{E^{(\tau)}}{\sigma^2}, \qquad \mathrm{MC}_{\text{tot}}=\sum_{\tau=1}^{\infty} C(\tau)\le n,$$9, while stability is measured through attractor counts, transient lengths, and the Derrida coefficient $n$0. High-synergy systems are reported as unstable and chaotic, with $n$1 and $n$2 bits; redundant systems are extremely stable, with $n$3 and $n$4 bits; TSE-complex systems lie in between, with $n$5 and $n$6 bits (Varley et al., 2024). The paper’s claim is direct: high-synergy systems are unstable and chaotic but have high information-integration capacity, whereas redundant systems are very stable but computationally impoverished.

A plausible implication is that several “edge-of-chaos” narratives in machine learning and complex systems are specific instances of the same structural pattern. The ESN result ties useful memory to $n$7 (Singh et al., 24 Jul 2025), while the Boolean-network study places TSE-complex systems between redundant order and synergistic chaos (Varley et al., 2024). The common theme is not maximal instability, but marginal stability that preserves distinguishability without collapsing trajectories too quickly.

5. Communication, storage infrastructure, and network control

Distributed storage systems make the trade-off fully operational. With $n$8 unreliable nodes of capacity $n$9, erasure rate $\Phi_\Theta(S,x)$0, and repair read rate $\Phi_\Theta(S,x)$1, the asymptotic source-data capacity is

$\Phi_\Theta(S,x)$2

The converse states that asymptotically no system can exceed this amount, and the liquid and advanced liquid repairers asymptotically achieve it. The threshold interpretation is sharp: if $\Phi_\Theta(S,x)$3, asymptotic capacity collapses to zero; increasing $\Phi_\Theta(S,x)$4 raises the storable fraction of $\Phi_\Theta(S,x)$5 by reducing the redundancy needed for long-term recoverability (Luby, 2016, Luby et al., 2021). Here, stability means continued recoverability under random failures, and capacity is exactly the usable payload that survives those failures.

In a stochastic fold model of traffic flow, the same phrase refers to a different mechanism. Capacity is the maximum stationary flow in the fundamental diagram, while stability is the asymptotic stability of free-flow and congested regimes. A deterministic fold model yields a continuous two-phase diagram without capacity drop. Introducing multiplicative stochastic noise in a single transition coefficient produces two opposing effects: it extends the stochastic stability region of the free-flow state beyond the deterministic fold point, but it also induces large variance in the congested regime and generates an inverse-$\Phi_\Theta(S,x)$6 capacity drop. The analysis identifies an effective stochastic fold $\Phi_\Theta(S,x)$7 and derives stationary mean and variance for the slow-state population when $\Phi_\Theta(S,x)$8 (Melo et al., 19 Mar 2025). This case is important because it reverses a common intuition: noise can increase free-flow stability while lowering effective congested capacity.

Cognitive radio systems express the trade-off between sensing quality and secondary throughput. In the model with steerable directional antennas, the secondary transmitter senses for $\Phi_\Theta(S,x)$9 seconds and then transmits during the remaining fraction $S$0 if the channel is sensed idle. The sensing probabilities

$S$1

improve with longer sensing time and with antenna orientation toward the primary transmitter, but both choices reduce secondary data rate or desired-link gain. The optimal transmit power is the minimum of a peak-power cap and an interference-outage cap,

$S$2

and the paper shows that the capacity curve in $S$3 has an interior maximizer $S$4 (Yazdani et al., 2018). In this formulation, stability is the protected operating condition of the primary user, enforced through an outage interference constraint.

Finite-regime network flow extends the idea to variable min-cuts. With link-rate intervals $S$5, network throughput is bounded by

$S$6

The paper defines stability as invariance of the min-cut set across all realizations. An unstable graph can have up to $S$7 distinct min-cut sets. At the same time, increasing the number of links on a cut reduces throughput variability substantially; the reported example shows nearly $S$8 reduction. To enforce stability, the paper proposes an algorithm of complexity $S$9 that shrinks uncertainty intervals so a single min-cut remains active, with the worst-case effect $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$0 (Gitik et al., 20 Jul 2025). The trade-off is therefore explicit: structural stability can be purchased by sacrificing best-case variability slack while preserving mean rates.

6. Electronic, thermodynamic, and adjacent formulations

At the device level, h-BN encapsulated GFETs exhibit the trade-off between RF capability and circuit stability. Capacity is the effective RF performance, measured by $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$1, $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$2, and power gain. With channel scaling to $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$3–$$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$4 nm, the projected extrinsic cutoff frequency approaches $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$5 THz, and for $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$6 nm the projected $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$7 reaches about $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$8 THz with $$C_{\text{source}} \approx \left(1-\frac{E}{2R}\right)Nc,$$9. Yet the same short-channel, high-gain regime expands negative-differential-resistance regions in which the Rollett factor can satisfy $E$0, making the device unstable at high frequency. Lower gate resistance simultaneously improves stability and raises $E$1, and the paper notes that high power gain and stability remain feasible even far from current saturation (Feijoo et al., 2018). The resulting trade-off is not between low and high gain alone, but between operating near the most aggressive NDR-enhanced gain region and maintaining a usable stability margin.

Information geometry provides a thermodynamic version of the concept. For a probability distribution $E$2, the intrinsic speed is

$E$3

and the paper proposes monotone decay of this quantity as a generalized stability criterion. It also derives the Cramér–Rao-type trade-off

$E$4

together with bounds linking excess entropy production to intrinsic speed. In this language, stability means decay of intrinsic speed near a steady state, while the capacity for rapid, precise observable change is bounded by available fluctuations and dissipation (Ito, 2019). The paper thereby interprets stability as an upper bound on local information-processing speed.

A related but not identical usage appears in other systems. In bosonic communication, trade-off coding allocates photons among classical, quantum, and entanglement resources, or among public, private, and secret-key resources; under high mean photon number the paper derives a rule of thumb $E$5, meaning that only an $E$6 number of photons need be devoted to quantum resources to operate near quantum capacity while the remaining power supports classical throughput (Wilde et al., 2011). In CLR-DRAM, the configurable constraint is capacity versus latency rather than dynamical stability: high-performance mode halves row capacity but reduces $E$7 by 60.1%, $E$8 by 64.2%, $E$9 by 46.4%, and $R$00 by 35.2%, exposing a runtime, row-granular capacity–latency control rather than a fixed design-time point (Luo et al., 2020).

Taken together, these results indicate that the capacity–stability trade-off is not a single theorem but a recurring structural motif. In some domains it is governed by contraction versus memory depth, in others by eigenvector misalignment, repair bandwidth, delay-dependent drift, disorder-induced phase diffusion, or interference protection. What remains invariant is the underlying logic: preserving reliable operation requires constraints that consume usable degrees of freedom, and relaxing those constraints expands capacity only by accepting greater sensitivity, shorter horizons, larger energy or repair costs, or a smaller robustness margin.